Benchmarking Generative Trajectory Models for Active-Inference Control

Yulin Li; Mohsen A. Jafari; Andrea Matta

cs.ROOct 5, 2026

Benchmarking Generative Trajectory Models for Active-Inference Control

Authors: Yulin Li, Mohsen A. Jafari, Andrea Matta

Organizations: Rutgers University–New Brunswick, Piscataway NJ 08854, USA · Politecnico di Milano, Via La Masa 1, 20156 Milano, Italy

Abstract

Learning from trajectory demonstrations offers a route to active-inference control of complex systems whose dynamics are difficult to model explicitly. We introduce generative active-inference control (GenAIF), in which one generative trajectory model learns from demonstrations and measured action interventions to supply a goal-conditioned policy distribution and a state-to-observation likelihood mapping. From this control design, we derive three model requirements: (i) useful action proposals, (ii) accurate prediction under imposed actions, and (iii) probabilistic observation evidence for belief updating and expected information gain. We benchmark diffusion, autoregressive Transformers, conditional variational autoencoders (CVAEs), and flow matching in a MuJoCo manipulation task with multiple physical conditions. Diffusion delivers the strongest control across the tested dynamics, while CVAE combines comparable short-horizon prediction with much faster inference. Correct conditioning is decisive, and trajectory reuse offers further computational savings. With the same frozen models, a hidden-dynamics experiment demonstrates prompt belief adaptation after an unannounced tilt change; subsequent instability identifies sustained inference as a remaining challenge. These findings support the use of shared generative trajectory models to connect action proposal, controlled prediction, and observation evidence within GenAIF.

Figures & tables

Figure 1: Ball-pushing task at illustrative lateral tilts. The pusher moves the ball around the obstacle toward the green goal; lines illustrate routes.

Figure 2: E1 control and prediction. (a) Control success with correct or incorrect tilt inputs (zero-lateral-tilt subset: 30 episodes; all tilts: 90 per group). (b) Mean Euclidean ball-position error of the predictive mean over six steps. Whiskers and shading show 95% bootstrap intervals over 10 sampling seeds (a) and 18 source trajectories (b); bands are pointwise. AR: autoregressive Transformer.

Model	Direct	Reuse	RMSE (mm)	Direct	Reuse
	Success /90		$H=6$ position	Time/action (ms)
Diffusion	81	68	14.295	57.123	9.794
Autoregressive	2	2	65.809	48.735	12.982
CVAE	49	61	15.264	5.643	1.236
Flow matching	43	34	24.776	199.571	37.037

Table 1: E1 control, prediction, and computation. Position RMSE uses 841 valid endpoints. Times cover model evaluation and checking per executed action, excluding simulator execution.

Model	Accuracy (%)	Log loss (nats)	Calibration error
Diffusion	98.96	0.540	0.062
Autoregressive	79.28	0.959	0.127
CVAE	99.07	0.434	0.057
Flow matching	99.07	0.474	0.069

Table 2: E1 hidden-tilt identification over 864 action–observation pairs. Lower log loss and calibration error indicate better probability estimates.

Figure 3: E2 belief tracking after a

0^{\circ}\!\to\!+15^{\circ}

tilt switch. Lines show mean

p(+15^{\circ})

across 12 trials; observation 11 first reflects the switch. Shading shows pointwise 95% bootstrap intervals over source episodes. The horizontal line marks the uniform prior.

Model	Identification delay median [min,max]	Identified by six	Correct at observation 30
Diffusion	3 [2,4]	12/12	5/12
Autoregressive	5 [5,8]	10/12	7/12
CVAE	4 [3,10]	10/12	0/12
Flow matching	4 [3,4]	12/12	0/12

Table 3: E2 adaptation across 12 switched trials. Delay counts observations after the switch until

p(+15^{\circ})

first reaches 0.9; final correctness uses the most likely hypothesis at observation 30.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Family	Success	Feasible (%)	Proposal (ms)	Plan (ms)
BC-MDN	11/18	82.4	0.75	346
CVAE	17/18	92.1	0.39	352
Diffusion	18/18	100.0	50.05	411
Flow	14/18	83.9	3.86	354
Autoregressive	14/18	82.1	14.03	360

Appendix

Table A1: Proposal-family comparison at

K=128

. Feasibility averages the fraction of feasible candidates per rollout; times are per decision.

Figure A1: Proposal-only control with MuJoCo transitions. Success versus (A) candidate budget, (B) measured planning time, and (C,D) requested CPU targets for nominal and new start–goal scenes. Diffusion exceeds the two smallest CPU targets, so pooled comparisons there are not at equal measured cost.

Method	$K$	Nominal	Transfer	Nominal	Transfer
		Success /18		Plan time (ms)
CEM	8	0	0	44.5	44.4
CVAE	16	16	18	43.8	43.5

Appendix

Table A2: CEM and CVAE at a 50 ms CPU target. Transfer uses new start–goal scenes; planning time includes proposal generation, rollout, and scoring.

Explore similar work

Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making
May 15, 2026Fan Feng, Selena Ge, Minghao Fu +6Diffusion Policies Latent Dynamics
University of California San Diego · MBZUAI · Carnegie Mellon University +2
Enfold: Folding World Model Imagination into Predictive Representations for Ultra-Efficient Embodied Control
Jul 29, 2026Weili Zeng, Yitong Xing, Fulong Liu +10Playable Video World Generation Goal-Conditioned Dynamic Manipulation
Shanghai Jiao Tong University · Qilu University of Technology (Shandong Academy of Sciences)
Renormalising Generative Models for Active Inference: Foundations, Derivations, and Verification
Aug 10, 2026Karim Zaghw, Andrew Pashea, Marc Pritsch +3Generative Models Derivations
University of Tübingen, Tübingen, Germany · Division of the Social Sciences, University of Chicago, Chicago, IL, USA · Heidelberg University, Heidelberg, Germany +3

May 15, 2026, cs.LGJ/K move · Enter open · S save

Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making

Fan Feng, Selena Ge, Minghao Fu, Zijian Li+5

University of California San Diego · MBZUAI · Carnegie Mellon University+2

Diffusion Policies Latent Dynamics Diffusion Models

Open paper AskPDF arXiv

Recent work has framed decision-making as a sequence modeling problem using generative models such as diffusion models. Although promising, these approaches often overlook latent factors that exhibit evolving dynamics, elements that are fundamental to environment transitions, reward structures, and high-level agent behavior. Explicitly modeling these hidden processes is essential for both precise dynamics modeling and effective decision-making. In this paper, we propose a unified framework that explicitly incorporates latent dynamic inference into generative decision-making from minimal yet sufficient observations. We theoretically show that under mild conditions, the latent process can be identified from small temporal blocks of observations. Building on this insight, we introduce Ada-Diffuser, a causal diffusion model that learns the temporal structure of observed interactions and the underlying latent dynamics simultaneously, and furthermore, leverages them for planning and control. With a modular design, Ada-Diffuser supports both planning and policy learning tasks, enabling adaptation to latent variations in dynamics, rewards, and latent actions. Experiments on simulated control and robotic benchmarks demonstrate its effectiveness in accurate latent inference and adaptive policy learning.

Abstract

Explore similar work

Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making

Enfold: Folding World Model Imagination into Predictive Representations for Ultra-Efficient Embodied Control

Renormalising Generative Models for Active Inference: Foundations, Derivations, and Verification